TL;DR. Best reward comes with the LoRA at 50% (0.623 vs 0.598 for the evolved prompt alone and 0.498 for BASE), so pair the evolved steering prompt with a 50% merge. The trade-off is intelligibility: emotion and blend edge up (blend 0.497, the highest of any condition) but quality dips (0.576 to 0.555) and WER climbs from 0.146 to 0.231. The main side-effect is a wistful, downcast color, LoRA50 correlates strongly with Longing (0.74), Sadness (0.73) and Pain (0.68) and pushes Emotional Numbness/Contentment up, so contemplation reads as calm and faintly melancholic. Do not push to 100/150%, WER doubles to ~0.52 and reward collapses to 0.41.
Without a LoRA
Best: evolved prompt, no LoRA → reward 0.60
GENERAL: A voice to the extreme expressing contemplation, a slow, thoughtful voice, deep in contemplation, musing and reflective, in every breath.
SCRIPT: (slowly, deep in contemplative thought) "<neutral sentence>" · temp 1.1 top_p 0.95 top_k 25
Best reward comes with the LoRA at 50% (0.623 vs 0.598 for the evolved prompt alone and 0.498 for BASE), so pair the evolved steering prompt with a 50% merge. The trade-off is intelligibility: emotion and blend edge up (blend 0.497, the highest of any condition) but quality dips (0.576 to 0.555) and WER climbs from 0.146 to 0.231. The main side-effect is a wistful, downcast color, LoRA50 correlates strongly with Longing (0.74), Sadness (0.73) and Pain (0.68) and pushes Emotional Numbness/Contentment up, so contemplation reads as calm and faintly melancholic. Do not push to 100/150%, WER doubles to ~0.52 and reward collapses to 0.41.